k-Best hidden Markov model decoding for unit selection in concatenative sound synthesis

Abstract

Concatenative synthesis is a sample-based approach to sound creation used frequently in speech synthesis and, increasingly, in musical contexts. Unit selection, a key component, is the process by which sounds are chosen from the corpus of samples. With their ability to match target units as well as preserve continuity, Hidden Markov Models are often chosen for this task, but one common criticism is its singular path output which is considered too restrictive when variations are desired. In this article, we propose considering the problem in terms of k-Best path solving for generating alternative lists of candidate solutions and summarise our implementations along with some practical examples.